Benchmark improvements (#109)

benchmark/benchmark_comparison_non_stream_WWR.jl

@@ -125,43 +125,43 @@ end
 rng = Xoshiro(42);
 rngs = Tuple(Xoshiro(rand(rng, 1:10000)) for _ in 1:Threads.nthreads());
 
-a = collect(1:10^7);
+a = collect(1:10^8);
 wsa = Float64.(a);
 
 times_other_parallel = Float64[]
-for i in 0:6
-    b = @benchmark sample_parallel_2_pass($rngs, $a, $wsa, 10^$i)
+for i in 0:7
+    b = @benchmark sample_parallel_2_pass($rngs, $a, $wsa, 10^$i) seconds=20
     push!(times_other_parallel, median(b.times)/10^6)
     println("other $(10^i): $(median(b.times)/10^6) ms")
 end
 
 times_other = Float64[]
-for i in 0:6
-    b = @benchmark sample($rng, $a, Weights($wsa), 10^$i; replace = true)
+for i in 0:7
+    b = @benchmark sample($rng, $a, Weights($wsa), 10^$i; replace = true) seconds=20
     push!(times_other, median(b.times)/10^6)
     println("other $(10^i): $(median(b.times)/10^6) ms")
 end
 
 ## single thread
 times_single_thread = Float64[]
-for i in 0:6
-    b = @benchmark weighted_reservoir_sample($rng, $a, $wsa, 10^$i)
+for i in 0:7
+    b = @benchmark weighted_reservoir_sample($rng, $a, $wsa, 10^$i) seconds=20
     push!(times_single_thread, median(b.times)/10^6)
     println("sequential $(10^i): $(median(b.times)/10^6) ms")
 end
 
 # multi thread 1 pass - 6 threads
 times_multi_thread = Float64[]
-for i in 0:6
-    b = @benchmark weighted_reservoir_sample_parallel_1_pass($rngs, $a, $wsa, 10^$i)
+for i in 0:7
+    b = @benchmark weighted_reservoir_sample_parallel_1_pass($rngs, $a, $wsa, 10^$i) seconds=20
     push!(times_multi_thread, median(b.times)/10^6)
     println("parallel $(10^i): $(median(b.times)/10^6) ms")
 end
 
 # multi thread 2 pass - 6 threads
 times_multi_thread_2 = Float64[]
-for i in 0:6
-    b = @benchmark weighted_reservoir_sample_parallel_2_pass($rngs, $a, $wsa, 10^$i)
+for i in 0:7
+    b = @benchmark weighted_reservoir_sample_parallel_2_pass($rngs, $a, $wsa, 10^$i) seconds=20
     push!(times_multi_thread_2, median(b.times)/10^6)
     println("parallel $(10^i): $(median(b.times)/10^6) ms")
 end
@@ -170,13 +170,13 @@ py"""
 import numpy as np
 import timeit
 
-a = np.arange(1, 10**7+1, dtype=np.int64);
-wsa = np.arange(1, 10**7+1, dtype=np.float64)
+a = np.arange(1, 10**8+1, dtype=np.int64);
+wsa = np.arange(1, 10**8+1, dtype=np.float64)
 p = wsa/np.sum(wsa);
 
 def sample_times_numpy():
     times_numpy = []
-    for i in range(7):
+    for i in range(8):
         ts = []
         for j in range(11):
             t = timeit.timeit("np.random.choice(a, size=10**i, replace=True, p=p)", 
@@ -196,20 +196,20 @@ ax1 = Axis(f[1, 1], yscale=log10, xscale=log10,
 	   yminorticksvisible = true, yminorgridvisible = true, 
 	   yminorticks = IntervalsBetween(10))
 
-scatterlines!(ax1, [10^i/10^7 for i in 1:6], times_numpy[2:end], label = "numpy.choice sequential", marker = :circle, markersize = 12, linestyle = :dot)
-scatterlines!(ax1, [10^i/10^7 for i in 1:6], times_other[2:end], label = "StatsBase.sample sequential", marker = :rect, markersize = 12, linestyle = :dot)
-scatterlines!(ax1, [10^i/10^7 for i in 1:6], times_other_parallel[2:end], label = "StatsBase.sample parallel (2 passes)", marker = :diamond, markersize = 12, linestyle = :dot)
-scatterlines!(ax1, [10^i/10^7 for i in 1:6], times_single_thread[2:end], label = "WRSWR-SKIP sequential", marker = :hexagon, markersize = 12, linestyle = :dot)
-scatterlines!(ax1, [10^i/10^7 for i in 1:6], times_multi_thread[2:end], label = "WRSWR-SKIP parallel (1 pass)", marker = :cross, markersize = 12, linestyle = :dot)
-scatterlines!(ax1, [10^i/10^7 for i in 1:6], times_multi_thread_2[2:end], label = "WRSWR-SKIP parallel (2 passes)", marker = :xcross, markersize = 12, linestyle = :dot)
-Legend(f[1,2], ax1, labelsize=10, framevisible = false)
+scatterlines!(ax1, [10^i/10^8 for i in 2:7], times_numpy[3:end], label = "numpy.choice sequential", marker = :circle, markersize = 12, linestyle = :dot)
+scatterlines!(ax1, [10^i/10^8 for i in 2:7], times_other[3:end], label = "StatsBase.sample sequential", marker = :rect, markersize = 12, linestyle = :dot)
+scatterlines!(ax1, [10^i/10^8 for i in 2:7], times_other_parallel[3:end], label = "StatsBase.sample parallel (2 passes)", marker = :diamond, markersize = 12, linestyle = :dot)
+scatterlines!(ax1, [10^i/10^8 for i in 2:7], times_single_thread[3:end], label = "WRSWR-SKIP sequential", marker = :hexagon, markersize = 12, linestyle = :dot)
+scatterlines!(ax1, [10^i/10^8 for i in 2:7], times_multi_thread[3:end], label = "WRSWR-SKIP parallel (1 pass)", marker = :cross, markersize = 12, linestyle = :dot)
+scatterlines!(ax1, [10^i/10^8 for i in 2:7], times_multi_thread_2[3:end], label = "WRSWR-SKIP parallel (2 passes)", marker = :xcross, markersize = 12, linestyle = :dot)
+Legend(f[2,1], ax1, labelsize=10, framevisible = false, orientation = :horizontal)
 
 ax1.xtickformat = x -> string.(round.(x.*100, digits=10)) .* "%"
 ax1.title = "Comparison between weighted sampling algorithms in a non-streaming context"
-ax1.xticks = [10^(i)/10^7 for i in 1:6]
+ax1.xticks = [10^(i)/10^8 for i in 2:7]
 
 ax1.xlabel = "sample ratio"
 ax1.ylabel = "time (ms)"
 
 f
-save("comparison_WRSWR_SKIP_alg.png", f)
+save("comparison_WRSWR_SKIP_alg_no_stream.png", f)

benchmark/benchmark_comparison_stream.jl

@@ -3,15 +3,15 @@ using Random, Printf, BenchmarkTools
 using CairoMakie
 
 rng = Xoshiro(42);
-stream = Iterators.filter(x -> x != 10, 1:10^7);
+stream = Iterators.filter(x -> x != 1, 1:10^8);
 pop = collect(stream);
 w(el) = Float64(el);
 weights = Weights(w.(stream));
 
 algs = (AlgL(), AlgRSWRSKIP(), AlgAExpJ(), AlgWRSWRSKIP());
 algsweighted = (AlgAExpJ(), AlgWRSWRSKIP());
 algsreplace = (AlgRSWRSKIP(), AlgWRSWRSKIP());
-sizes = (10^3, 10^4, 10^5, 10^6)
+sizes = (10^4, 10^5, 10^6, 10^7)
 
 p = Dict((0, 0) => 1, (0, 1) => 2, (1, 0) => 3, (1, 1) => 4);
 m_times = Matrix{Vector{Float64}}(undef, (3, 4));
@@ -24,13 +24,13 @@ for m in algs
         replace = m in algsreplace
         weighted = m in algsweighted
         if weighted
-            b1 = @benchmark itsample($rng, $stream, $w, $size, $m) evals=1
-            b2 = @benchmark sample($rng, collect($stream), Weights($w.($stream)), $size; replace = $replace) evals=1
-            b3 = @benchmark sample($rng, $pop, $weights, $size; replace = $replace) evals=1
+            b1 = @benchmark itsample($rng, $stream, $w, $size, $m) seconds=20
+            b2 = @benchmark sample($rng, collect($stream), Weights($w.($stream)), $size; replace = $replace) seconds=20
+            b3 = @benchmark sample($rng, $pop, $weights, $size; replace = $replace) seconds=20
         else
-            b1 = @benchmark itsample($rng, $stream, $size, $m) evals=1
-            b2 = @benchmark sample($rng, collect($stream), $size; replace = $replace) evals=1
-            b3 = @benchmark sample($rng, $pop, $size; replace = $replace) evals=1
+            b1 = @benchmark itsample($rng, $stream, $size, $m) evals=1 seconds=20
+            b2 = @benchmark sample($rng, collect($stream), $size; replace = $replace) seconds=20
+            b3 = @benchmark sample($rng, $pop, $size; replace = $replace) seconds=20
         end
         ts = [median(b1.times), median(b2.times), median(b3.times)] .* 1e-6
         ms = [b1.memory, b2.memory, b3.memory] .* 1e-6
@@ -39,6 +39,7 @@ for m in algs
             push!(m_times[r, c], ts[r])
             push!(m_mems[r, c], ms[r])
         end
+        println("c")
     end
 end
 

benchmark/benchmark_comparison_stream_WWR.jl

@@ -65,14 +65,61 @@ a = Iterators.filter(x -> x != 1, 1:10^8)
 wv_const(x) = 1.0
 wv_incr(x) = Float64(x)
 wv_decr(x) = 1/x
-wvs = (wv_decr, wv_const, wv_incr)
+wvs = ((:wv_decr, wv_decr), 
+       (:wv_const, wv_const),
+       (:wv_incr, wv_incr))
 
-for wv in wvs
-    for m in (AlgWRSWRSKIP(), AlgAExpJWR())
-        for sz in [10^i for i in 0:7]
-            b = @benchmark itsample($a, $wv, $sz, $m) seconds=10
-            println(wv, " ", m, " ", sz, " ", median(b.times))
+benchs = []
+for (wvn, wv) in wvs
+    for m in (AlgAExpJWR(), AlgWRSWRSKIP())
+        bs = []
+        for sz in [10^i for i in 3:7]
+            b = @benchmark itsample($a, $wv, $sz, $m) seconds=20
+            push!(bs, median(b.times))
+            println(median(b.times))
         end
+        push!(benchs, (wvn, m, bs))
+        println(benchs)
     end
 end
 
+using CairoMakie
+
+f = Figure(backgroundcolor = RGBf(0.98, 0.98, 0.98), size = (1100, 700));
+
+f.title = "Comparison between AExpJ-WR and WRSWR-SKIP Algorithms"
+
+ax1 = Axis(f[1, 1], yscale=log10, xscale=log10, 
+       yminorticksvisible = true, yminorgridvisible = true, 
+       yminorticks = IntervalsBetween(10))
+ax2 = Axis(f[1, 2], yscale=log10, xscale=log10, 
+       yminorticksvisible = true, yminorgridvisible = true, 
+       yminorticks = IntervalsBetween(10))
+ax3 = Axis(f[1, 3], yscale=log10, xscale=log10, 
+       yminorticksvisible = true, yminorgridvisible = true, 
+       yminorticks = IntervalsBetween(10))
+
+#ax4 = Axis(f[2, 1])
+
+for x in benchs
+    label = x[1] == :wv_const ? (x[2] == AlgAExpJWR() ? "ExpJ-WR" : "WRSWR-SKIP") : ""
+    ax = x[1] == :wv_decr ? ax1 : (x[1] == :wv_const ? ax2 : ax3)
+    marker = x[2] == AlgAExpJWR() ? :circle : (:xcross)
+    scatterlines!(ax, [10^i/10^8 for i in 3:7], x[3] ./ 10^6, marker = marker, 
+                  label = label, markersize = 12, linestyle = :dot)
+end
+
+Legend(ax4, labelsize=10, framevisible = false, orientation = :horizontal)
+
+for ax in [ax1, ax2, ax3]
+    ax.xtickformat = x -> string.(round.(x.*100, digits=10)) .* "%"
+    #ax.ytickformat = y -> y .* "^"
+    ax.title = ax == ax1 ? "decreasing weights" : (ax == ax2 ? "constant weights" : "increasing weights")
+    ax.xticks = [10^(i)/10^8 for i in 3:7]
+    ax.yticks = [10^i for i in 2:4]
+    ax.xlabel = "sample ratio"
+    ax == ax1 && (ax.ylabel = "time (ms)")
+end
+
+save("comparison_WRSWR_SKIP_alg_stream.png", f)
+f

src/SamplingUtils.jl

@@ -34,6 +34,7 @@ Base.eltype(::SeqSampleIterWR) = Int
 Base.IteratorSize(::SeqSampleIterWR) = Base.HasLength()
 Base.length(s::SeqSampleIterWR) = s.n
 
+# courtesy of StatsBase.jl for part of the implementation
 struct SeqSampleIter{R}
     rng::R
     N::Int